WIP: Human-ai interactions in real-world complex environments using a comprehensive reinforcement learning framework

摘要: Deep reinforcement learning (RL) has successfully tackled many real-world tasks. However, these algorithms suffer from the wellknown sample-inefficiency problem. Deep RL systems usually require millions of environment interactions to learn and have stable performance. In this work, we show that human-AI teams outperform human-only controlled and fully autonomous teams for complex tasks. We develop a novel simulator for a critical infrastructure scenario and a user interface for humans to effectively advise AI agents. We show that humans can provide useful advice to the RL agents, allowing them to improve learning in a multi-agent setting.